MMK-DAR: A KAN hybrid expert-based approach for food risk prediction

Food safety is closely related to people's safety and health, economic development and social stability. However, how to accomplish rapid and accurate detection and control of pesticide residues, heavy metals and other risky substances at low concentrations is an urgent problem. In this paper, a novel multivariate time series prediction model, MMK-DAR (Multilayer Hybrid Kolmogorov-Arnold Network combined with Dynamic Autoregressive Correction), is proposed to enhance the performance of the existing risk detection methods for food products with the need of model interpretability in some aspects. The model is designed to significantly improve the performance and interpretability of multivariate time series prediction through the synergistic design of dynamic convolutional gating and autoregressive correction modules. The dynamic convolutional gating module utilizes multi-scale convolution to capture local time-series patterns, replacing the weight assignment mechanism of traditional linear gating and enhancing the model's ability to model variable-specific dynamics and local time dependence. The autoregressive correction module combines causal convolution and the attention mechanism, which is able to capture both short-term and long-term dependencies in the time series, further improving the accuracy and stability of the predictions. Experimental results on six mainstream time-series datasets as well as the camel's milk dataset show that MMK-DAR exhibits significant performance advantages in long-term time-series prediction tasks, validating its generalization ability in complex realistic scenarios.